Iris
The first intelligent system we built is our own. You can talk to it right now, while you read how it works.
The problem with selling intelligent systems
A studio that says it designs intelligent systems has a problem no presentation solves: anyone can say it. Decks all look alike, case studies are written after the fact, and the person listening has no way to verify anything before signing.
We decided not to tell you. Iris is a system running on this site, in text and in voice. Anyone weighing up working with us can test it before even writing to us — which is exactly what we would like them to do.
You don’t have to trust us. You can check.
Ask it which moment your brand is in. Ask it something off-topic, to see how it reacts. Ask it for free consulting and watch what happens: the refusal is as designed as the answer.
The decisions, not the technology
The language model is the least interesting part: the good ones are available to everyone. What separates one system from another are the rules you give it — and above all the ones you take away. These are ours, in force right now.
It does not give free consulting
The most common question is “how do you do it”. Iris does not explain how. It recognises the real problem underneath the question, names the brand moment you are in, and asks something that moves the conversation forward. A list of generic advice has never built a brand: giving it away devalues the work and helps nobody.
One question at a time
No questionnaire, no data capture dressed up as conversation. Two questions across the whole exchange, not five. It should feel like being accompanied, not examined.
Four sentences, never more
No bullet points, no bold, no emoji. If an answer needs more room, that is material for a Discovery Session, not for a chat. The length limit is the single rule that most makes it read like a person writing.
Two ways forward, never one
When it suggests continuing, it always offers both: book a call, or write through the form without booking anything. Not everyone wants to talk on the phone, and treating the form as the fallback loses precisely the most cautious people.
It does not pretend to be human
Ask whether you are talking to a person and it says it is Solun’s digital intelligence. No hedging, no drama. Pretending costs more, once discovered, than it ever earned.
It would rather admit it doesn’t know
It does not invent clients, numbers, case studies or skills we don’t have. When it has no answer it says so and offers to look into it with a person. That is the hardest rule to enforce on a language model, and the one we spent the most time on.
A system, not a widget
Two channels, one voice
Text chat and real-time voice conversation, with the same personality and the same rules. Switching channel should not mean talking to someone else.
The personality is a document, not code
How Iris speaks is written in plain language and versioned: it changes from the admin panel without touching a line of code, and every previous version stays readable. A brand that shifts its tone should not have to wait for a developer.
Every conversation can be reviewed
We read the exchanges back to correct wrong answers and understand what people actually ask. The questions left unanswered are the most useful material we have for improving the site, not just the assistant.
Italian and English
It answers in the language you write in, with the same care in both. Not a machine translation of something conceived in another language.
How we know it is working well
An assistant that answers is not the same as an assistant that answers well. And the difference is invisible: an agent does not break with an error, it degrades quietly. The model changes, the knowledge base is updated, a question nobody anticipated arrives — and three months later it answers worse than it started, without anyone noticing.
It seemed like a question we ought to be able to answer about ourselves, before putting it to anyone else. So we built an internal system for evaluating conversational agents.
Two ways of looking
Some things are simply measured: an empty answer, an answer that ignores the knowledge available, a tool that fails while the agent claims it succeeded, a guardrail worked around. They either pass or they do not, and there is no interpretation. The rest calls for judgement — whether the tone is the brand’s, whether the conversation was handled well, whether an answer asserts things nobody told it — and on those a model does the assessing, not a rule.
Who checks the checker
An automated evaluator can be as wrong as the thing it evaluates, and a system that awards itself a good mark proves nothing. So we compare the automated judgements with those of people, who assess the same conversations without seeing the machine’s verdict. Where the two diverge, the error may lie on either side — and it matters which. It is the slowest part of the work, and the part that makes the rest worth anything.
- Evaluate
- Compare with human judgement
- Set a baseline
- Correct
- Re-evaluate
It is not a one-off test: after the last step it begins again.
Where it stands
It is an internal project, in validation: not a product for sale, and we do not present it as a finished platform. We do not publish Iris’s scores, because a published number becomes a promise, and promises about conversational systems age badly. We would rather tell you how we measure it, and let Iris answer you.
Where it is going
Today Iris orients: it works out where you are and points you in a direction. The next step is the document base — the programme texts, the method, the questions we get most often — so it can answer on substance and cite where the information came from.
We write this here because it is where the difference between a real system and a demo shows: a real system has a next version, and the people who built it can tell you what it is.
What Iris does not do
It does not decide for you, does not give legal or tax advice, does not replace talking to a person, and like every system of this kind it can be wrong. We wrote a whole page about its limits, because stating them is part of the design.
Read the AI transparency pageIf you need a system like this
Iris is the demonstration of how we work: first we decide what the system must do and what it must not do, then we build it. If you are considering something similar for your company, the first step is a conversation.
Book a Discovery Session